ECCV 2026 Papers

3D object detection is a core task in autonomous driving per-ception, with mainstream methods primarily based on LiDAR. However,its robustness deteriorates significantly in adverse weather conditions.While 4D radar offers all-weather robustness and unique Doppler veloc-ity, it is hindered by its inherent sparsity and noise interference. There-fore, fusing LiDAR and 4D radar is highly promising. However, due to thesignificant differences in semantics and density between the data of thesetwo modalities, single feature space fusion cannot fully leverage the ad-vantages of each modality and the complementarity between modalities.To address this, we propose a two-stage fusion detection framework—HVA-Fusion. Specifically, motion velocity is first estimated through aMotion Velocity Estimation and Encoding module to enhance dynamicperception. Then, in the pillar feature space, a Cross-Modal Progres-sive Adaptation module is introduced for local feature alignment andfusion (Stage-I fusion) to enrich representations and mitigate semanticdifferences between modalities. After that, to alleviate the sparsity of4D radar point cloud, a Radar Cross-Section (RCS) Guided GaussianGeneration module is developed to densify the feature representations.Subsequently, in the Bird’s Eye View (BEV) space, the enhanced multi-modal features are adaptively fused through a Multi-Scale Bi-DirectionalDeformable Attention Gate module (Stage-II fusion) to address featuremisalignment and sensor degradation caused by adverse weather. Exten-sive experiments demonstrate that HVA-Fusion achieves highly compet-itive performance across the K-Radar, VoD, and TJ4DRadSet datasets,while exhibiting robustness in adverse weather conditions.

Federated learning (FL) enables collaborative model trainingwithout sharing raw data, but practical deployments must balance sta-tistical heterogeneity, resource-constrained clients, and communicationefficiency. While conventional FL requires clients to train full modelslocally, split federated learning (SFL) reduces client-side computationby partitioning models across tiers, at the cost of transmitting high-dimensional activations and gradients during training. We propose Hier-archical Split-Federated Prototyping (H-SFP), a communication-efficientframework for hierarchical cloud–edge–client environments. Instead ofexchanging sample-level activations or gradients, H-SFP communicatescompact class-wise feature statistics and uses them to synthesize repre-sentative feature distributions at upper tiers. This statistical prototyp-ing mechanism enables independent training of model segments whilepreserving the computational advantages of split-model execution andsubstantially reducing synchronization and communication overhead. Akey finding of this work is that lightweight first- and second-order fea-ture statistics are sufficient to support hierarchical split-model train-ing, providing an effective communication interface without requiringsample-level feature exchange. We analyze the communication proper-ties and stability of the proposed statistical interface under hierarchi-cal aggregation. Extensive experiments on CIFAR-10/100, HAM10000,ImageNet-1K, and ISIC-2018 demonstrate that H-SFP remains robustunder heterogeneous data distributions while reducing communicationoverhead by up to two orders of magnitude compared with federatedand split-learning baselines.

Generating complete 3D objects under partial occlusions (i.e.,amodal scenarios) is a practically important yet challenging problem,as large portions of object geometry are unobserved in real-world sce-narios. Existing approaches either operate directly in 3D, which en-sures geometric consistency but often lacks generative expressiveness,or rely on 2D amodal completion, which provides strong appearancepriors but does not guarantee reliable 3D structure. This raises a keyquestion: how can we achieve both generative plausibility and geometriccoherence in amodal 3D modeling? To answer this question, we intro-duce GENA3D (GENerative Amodal 3D), a framework that integrateslearned 2D generative priors with explicit 3D geometric reasoning withina conditional 3D generation paradigm. The 2D priors enable the modelto plausibly infer diverse occluded content, while the 3D representationenforces multi-view consistency and spatial validity. Our design incorpo-rates a novel View-Wise Cross-Attention for multi-view alignment and aStereo-Conditioned Cross-Attention to anchor generative predictions in3D relationships. By combining generative imagination with structuralconstraints, GENA3D generates complete and coherent 3D objects fromlimited observations without sacrificing geometric fidelity. Experimentsdemonstrate that our method outperforms existing approaches in bothsynthetic and real-world amodal scenarios, highlighting the effectivenessof bridging 2D priors and 3D coherence in generating plausible and geo-metrically consistent 3D structures in complex environments.

Recent advances in Gaussian Splatting-based inverse rendering extend Gaussian primitives with shading parameters and physically grounded light transport, enabling high-quality material recovery from dense multi-view captures. However, the accuracy of these methods degrades under sparse-view settings, where limited observations lead to severe ambiguity between geometry, reflectance, and lighting. We introduce GAINS (Gaussian-based Inverse rendering from Sparse multi-view captures), a two-stage inverse rendering framework that leverages foundation models as priors to stabilize geometry and material estimation. The core technical contribution of this paper is an inverse rendering framework that unifies foundation model priors with physically-based representations in an optimization scheme. GAINS first refines geometry using monocular depth, normal, and diffusion priors, and then employs segmentation, intrinsic image decomposition (IID), and diffusion priors to regularize material recovery. Extensive experiments on synthetic and real-world datasets show that GAINS significantly improves material parameter accuracy, relighting quality, and novel-view synthesis compared to state-of-the-art Gaussian-based inverse rendering methods. While GAINS outperforms and remains competitive across a wide range of objects captured with 4 to 32 cameras, the improvement is particularly pronounced under sparse-view settings, where ambiguity is high and learning-based priors become especially beneficial. Project page: https://patrickbail.github.io/gains/

Zero-shot object counting aims to estimate the number ofobjects specified by a text query without category-specific training. Re-cent approaches primarily rely on density regression or detection-styleinstance prediction. While effective, density-based models often sufferfrom spatial ambiguity and background leakage due to weakly regulatedmass allocation, leading to fragmented or part-biased representationsthat increase counting error in complex scenes. In this work, we proposean instance-aware dual-decoder framework that structurally couples den-sity and point representations for zero-shot object counting. Instead oftreating density estimation as independent pixel-wise regression, we in-terpret it as a structured mass allocation problem over a latent set ofobject instances. Predicted instance centers induce a soft instance-wisedecomposition of the density map, upon which we enforce two geometricconstraints: (1) per-instance mass conservation, ensuring each object con-tributes approximately one unit of density mass, and (2) center-of-massalignment, encouraging each density component to concentrate aroundits corresponding predicted center. These constraints introduce instance-level geometric consistency and lead to more accurate mass allocation,thereby reducing counting error. Extensive experiments on FSC-147,PUCPR+, and CARPK show that our approach consistently reducescounting error and establishes new state-of-the-art performance in zero-shot object counting.

Human driving behavior is inherently diverse, yet most end-to-end autonomous driving (E2E-AD) systems learn a single averagedriving style, neglecting individual differences. Achieving personalizedE2E-AD faces challenges across three levels: limited real-world datasetswith individual-level annotations, a lack of quantitative metrics for eval-uating personal driving styles, and the absence of algorithms that canlearn stylized representations from users’ trajectories. To address thesegaps, we propose Person2Drive, a comprehensive personalized E2E-ADplatform and benchmark. It includes an open-source, flexible data col-lection system that simulates realistic scenarios to generate scalable, di-verse personalized driving datasets; style vector–based evaluation metricswith Maximum Mean Discrepancy and KL divergence to comprehen-sively quantify individual driving behaviors; and a personalized E2E-ADframework with a style reward model that efficiently adapts E2E mod-els for safe and individualized driving. Crucially, our framework enablesplug-and-play personalization by fine-tuning only the trajectory predic-tion head, preserving the pretrained base model and ensuring safety. Ex-tensive experiments demonstrate that Person2Drive enables fine-grainedanalysis and effective personalization, while preserving driving perfor-mance and success rate even in challenging scenarios.

Recent advances in Large Vision Language Models (LVLMs)have led to interest in alignment methods such as Direct PreferenceOptimization (DPO), which improves multi-modal instruction followingwithout costly supervision or explicit reward modeling. However, pre-vious attempts to improve DPO through data augmentation and pref-erence ranking fall short of bridging vision-language perceptual gaps,limiting consistent multi-modal alignment. In this paper, we proposea novel Vision-Language Consistency based Direct PreferenceOptimization (VLC-DPO), which automatically constructs prefer-ence data and leverages consistent information from visual inputs andcorresponding textual descriptions as alignment signals. The VLC-DPOconsists of three key components: (1) an initial response generation thatproduces image- and description-based responses to bridge the modal-ity gap in LVLMs by capturing complementary aspects of visual un-derstanding, (2) a preference data construction via vision-language con-sistency score that automatically selects high-quality preference pairsbased on response consistency, and (3) a VLC-DPO training that usesconsistency-based implicit rewards to optimize preferences, assigning dy-namic weights to less image-related or distant samples to improve DPOlearning. Extensive experiments on hallucination mitigation and zero-shot VQA demonstrate the e↵ectiveness of our method, which outper-forms existing approaches.

Reflection removal aims to recover clean images from pho-tographs degraded by glass, a crucial problem in mobile photography.Single-image methods suffer from the fundamental ill-posedness of de-composition and often rely on specialized hardware to enhance robust-ness. We propose a user-friendly dual-view diffusion framework that ef-fectively leverages complementary viewpoint information for ubiquitousdual-camera devices. Our approach introduces frame-aware spatial at-tention to capture implicit cross-view correlations and integrates camerapose as conditional embeddings for geometric guidance. We contributeDualRef, a dataset for reflection removal comprising semi-syntheticpairs and real-world dual-camera sequences with ground truth. Exper-iments demonstrate over 10% PSNR improvement on real-world data,significantly outperforming state-of-the-art methods.

Knowledge-Intensive Visual Question Answering (KI-VQA)benchmarks evaluate Vision–Language Models (VLMs) as multimodalknowledge assistants by requiring external information beyond a pro-vided image to answer questions. KI-VQA involves multiple sub-problems—referring expression understanding, visual grounding, object recogni-tion, knowledge retrieval, and reasoning—yet existing benchmarks typ-ically report only end-task accuracy, obscuring where failures arise. Toanalyze the full KI-VQA pipeline, we introduce CRAG-MM-Diagnostics,a diagnostic benchmark with stage-wise data annotations that isolate1 language-based visual grounding, 2 object identification, and 3knowledge retrieval and reasoning. We evaluate fully parametric andretrieval-augmented VLMs, providing fine-grained analyses using newlycollected metadata, such as target ROIs, entity names, and visual com-plexity scores. Our results point to knowledge retrieval and reasoning asthe primary bottleneck, but also highlight issues in the other parts ofthe KI-VQA pipeline, such as the fact that VLMs struggle with targetobject identification or that image retrievers struggle to integrate textualcues. These findings expose fundamental limitations in current KI-VQAsystems and motivate stage-aware evaluation. We, lastly, leverage thesefindings to propose a grounded bimodal RAG pipeline that integrates avisual grounding module to crop targets before image retrieval, boostingGPT-5 and Qwen’s respective accuracies by 13.3 and 8.5 percentage points.

Camouflaged Object Detection (COD) remains challengingdue to the deliberate alignment of foreground statistics with surround-ing background patterns, which induces strong scene-dependent ambi-guity. Existing approaches typically adopt object-centric modeling orreference-based augmentation. They often treat object appearance andbackground statistics independently, without explicitly modeling theirconditional relationship. This work treats COD as a scene-conditionedpattern-deviation reasoning problem and captures how a camouflagedobject deviates from its background scene. In this direction, we proposea pattern-aware scene-conditioned reasoning (PASR) method with twostages. The first stage constructs a scene-aligned background anchor viabackground-first retrieval, while the second stage performs a conditionedpatch-level deviation reasoning. A reference prototype library is con-structed in an offline and annotation-free manner, where foreground pro-totypes are derived from self-inferred rough masks rather than manuallycurated annotations. Unlike retraining-based paradigms, PASR achievesgeneralization through prototype-level expansion rather than parameteradaptation, enabling scalable inference across diverse background dis-tributions. The resulting deviation maps can serve as priors for genericsegmentation models. Extensive experiments on challenging COD bench-marks demonstrate that PASR outperforms existing unsupervised andweakly supervised methods and significantly narrows the gap with fullysupervised models.

Customized image editing equips diffusion models with spe-cific visual transformations using limited paired data, typically via Low-Rank Adaptation (LoRA). However, combining acceleration and effectLoRAs during inference often introduces interference, causing conceptbleeding, semantic drift, and degraded stylistic fidelity. In this paper, wepropose a unified paradigm that distills numerous customized conceptsand fast-inference capabilities into a single LoRA, effectively resolvingcompositional conflicts while reducing deployment overhead. To addressthe challenges of multi-concept distillation, particularly gradient van-ishing and training instability, we introduce CollectionLoRA, a multi-teacher hybrid distillation framework. Our framework incorporates: (1)Probabilistic Dual-Stream Routing (PDSR) that leverages general-domaindata as regularization to enhance robustness. (2) Automated Asymmet-ric Conditioning (AAC) utilizing Vision-Language Models (VLMs) andorthogonal trigger words to ensure concept isolation, and (3) the De-coupled Hybrid Objective (DHO) that synergizes Trajectory-AnchoredFlow Matching (TA-FM) priors with target simulation and distribu-tion matching, ensuring robust optimization across heterogeneous multi-task distributions. Extensive experiments demonstrate that our unifiedarchitecture mitigates multi-LoRA interference and achieves enhancedconcept fidelity, effective feature isolation, and high-quality rapid syn-thesis. Code is available at https://github.com/Qwen-Applications/CollectionLoRA.

Sim-to-real transfer remains a major obstacle for reinforce-ment learning (RL), especially for vision-based control where image ob-servations exacerbate the state-distribution shift between simulation andthe real world. Domain adaptation (DA) is a promising remedy for thischallenge. Prior sim-to-real DA works have demonstrated encouragingresults, yet these approaches typically assume substantially more targetdata, which is not available in practice. Indeed, their performance de-grades significantly when the target data budget is reduced. To addressthis challenge, we propose AIDA (Adaptive Imagination for DomainAdaptation), a domain adaptation framework for visual reinforcementlearning that addresses sim-to-real transfer under scarce target datawithout requiring additional interaction with the target environment.Our key idea is adaptive imagination: generating reliable and semanticimagination rollouts to augment limited target data. Specifically, AIDAemploys a distribution-shift-aware discriminator that truncates rolloutswhen imagined transitions drift into low-confidence regions, so that onlyreliable transitions contribute to the augmentation. On these reliabletransitions, AIDA introduces a self-consistency loss that cycles throughstate → image observation → state, penalizing discrepancies betweenthe original and reconstructed states. This provides additional adap-tation signals beyond the scarce target data. Our experiments demon-strate that adaptive imagination effectively truncates unreliable rollouts.By enforcing a self-consistency loss on the resulting reliable transitions,AIDA learns semantically meaningful state representations and outper-forms baselines across five MuJoCo tasks and two Gymnasium-Roboticstasks.

Camera pose matters. The position and orientation of eachviewpoint define a shared spatial coordinate frame that relates observa-tions across video frames. Yet this signal is absent from multimodal LLMs(MLLMs) for video understanding, which process frames as isolated 2Dsnapshots, instead of the persistent scene humans perceive. We revisitpose as a lightweight supervisory signal and introduce Cambrian-P , avideo MLLM augmented with per-frame learnable camera tokens and apose regression head. With a carefully designed sampling scheme, themodel achieves substantial gains of 4.5–6.5% on spatial reasoning bench-marks such as VSI-Bench, generalizes across eight additional spatial andgeneral video QA benchmarks, and, as a byproduct, achieves state ofthe art streaming pose estimation on ScanNet. Surprisingly, training onpseudo-annotated poses from in-the-wild video further improves generalvideo QA benchmarks, showing pose helps beyond spatial reasoning. To-gether, these results position camera pose as a fundamental signal forvideo models that reason about the physical world.

Universal photometric stereo (UniPS) aims to recover sur-face normals from images captured under unconstrained illumination,which fundamentally differs from calibrated photometric stereo (CPS),where lighting conditions are known. This gap forces most existing UniPSmethods to adopt fully end-to-end learning that struggles to disentangleillumination and surface geometry due to the lack of explicit physical con-straints. In this paper, we propose Bridge-UniPS, a novel frameworkthat bridges CPS and UniPS through a lighting adapter trained withoutdirect intermediate supervision. The adapter transforms images capturedunder unconstrained illumination into a sequence of image–illuminationpairs directly compatible with CPS models, allowing CPS to act as aphysics-aligned intermediate representation and thereby improving theaccuracy of surface normal recovery. The lighting adapter receives nodirect supervision on Bridge Images or target lighting directions, butis optimized through the final normal-estimation objective with gra-dients propagated through a frozen CPS network. It can still produceCPS-compatible Bridge Images while preserving illumination-geometryconsistency. This behavior suggests that modern learning-based CPSmodels provide useful physical priors to guide the reconstruction. Ex-tensive experiments on multiple benchmarks demonstrate that Bridge-UniPS achieves state-of-the-art performance under unconstrained illu-mination, exhibiting strong generalization and significantly improvingfine-scale surface normal accuracy.

Interactive 3D segmentation aims to extract object masks inpoint clouds with minimal user clicks. Despite recent progress, most ex-isting approaches still struggle with (i) coarse voxel resolution that blursfine boundaries under limited clicks and (ii) hard false positives causedby confusing background structures. These issues are exacerbated bydensity and scale shifts across datasets (e.g., dense RGB-D reconstruc-tions vs. sparse LiDAR scans), where fixed refinement heuristics andpurely click-driven decoding generalize poorly. To address them, we pro-pose NegROI — a novel transformer-based interactive framework thatcouples click-centric multi-resolution refinement with scene-conditionednegative prompts. Given a coarse voxel prediction, it refines only a lo-cal Region Of Interest (ROI) around the current click on a finer gridand fuses refined logits back to the coarse mask. To improve robustnessand efficiency, we introduce uncertainty-driven selective refinement thatprioritizes ambiguous regions. Meanwhile, we model hard backgroundpatterns via a set of scene-conditioned negative prompts obtained bycross-attention over scene tokens. We further stabilize these prompts witha diversity regularizer. Finally, we propose boundary-aware hard nega-tive mining to supervise negative-prompt attention toward boundary-proximal, high-confidence false positives. Our experiments on commonbenchmark datasets (i.e., ScanNet, S3DIS, and KITTI) demonstrate im-proved click efficiency and reduced false positives, with stronger cross-dataset robustness than the state-of-the-art baselines.

Multi-site functional MRI (fMRI) studies are essential forrobust neuropsychiatric diagnosis yet suffer severe domain shifts fromscanner heterogeneity, demographics, and site-specific acquisition pro-tocols. Traditional domain adaptation requires concurrent source andtarget data access, violating clinical privacy regulations. Moreover, func-tional connectivity matrices lie on the Symmetric Positive Definite (SPD)manifold, where Euclidean operations cause geometric distortions cor-rupting diagnostic patterns. We propose BrainRiem, a source-free do-main adaptation framework learning compact Riemannian brain pro-totypes via manifold-aware bi-level optimization. It employs the Log-Euclidean Metric to ensure prototypes remain valid SPD matrices, whileDirichlet Energy spectral calibration aligns their frequency characteris-tics with real brain networks. Only anonymized prototypes are trans-mitted to target sites, serving as stable anchors for training local mod-els without source data access and reducing leakage under the evalu-ated attacks. Comprehensive experiments on ABIDE and REST-meta-MDD show BrainRiem consistently outperforms state-of-the-art source-free, traditional, and graph domain adaptation methods across diversescanners and demographics. Notably, learned prototypes exhibit biolog-ically interpretable connectivity patterns aligning with established neu-roscience findings, validating the necessity of Riemannian geometry forbrain network analysis.

Existing label-deficient concealed object segmentation (LDCOS) methods either rely on consistency constraints within a meanteacher framework or employ the Segment Anything Model (SAM) as a fixed pseudo-label generator. Interestingly, task-specific segmenters and foundation models exhibit complementary failure modes: segmenters lack generalization under scarce labels, while SAM suffers from domain gaps in concealed scenes. A natural solution is to employ co-training to mutually promote both models. However, we observe that standard homogeneous co-training cannot exploit this complementarity because architecturally identical networks tend to share similar error patterns, especially when targets are heavily concealed. To address this, we present SCALER (Segmenter-Coordinated Asymmetric LEaRning), a framework that jointly optimizes a mean-teacher segmenter and a learnable SAM through two alternating phases with model-specific optimization strategies. In Phase I, the segmenter is updated under fixed SAM supervision using entropy-based image-level and uncertainty-based pixellevel weighting to suppress unreliable pseudo-label regions. In Phase II, SAM is updated via an augmentation invariance loss and a noise resistance loss, which exploit SAM’s inherent perturbation robustness rather than treating it identically to the segmenter. This asymmetric design is the key distinction from prior co-training and one-way distillation methods: each model’s learning strategy is tailored to its own inductive bias, enabling mutual enhancement. Experiments across eight LDCOS tasks demonstrate consistent gains, and SCALER improves both the lightweight segmenter and the foundation model, serving as a general plug-and-play paradigm for label-scarce conditions. The code is available at https://github.com/ChunmingHe/SCALER.

Real-world dynamics are inherently compositional: multiple entities move simultaneously within a shared scene, each exhibiting distinct motion patterns. Yet most existing video representations encode motion globally, without explicitly capturing localized motion for individual entities. Crucially, motion is defined relative to a global reference frame, including camera motion and scene layout. However, localized embeddings are often computed from cropped images or obtained by masking features after encoding, discarding the context needed to interpret motion. To address this, we introduce a promptable localized motion representation that produces persistent embeddings for user-specified regions defined by spatial masks. Rather than cropping the input or masking features, our model processes the full video and conditions motion encoding directly on the queried region. This yields temporally consistent, region-addressable embeddings that isolate local dynamics while retaining the global context required for disambiguation. We demonstrate object-level motion transfer, enabling controlled composition of dynamic scenes. Beyond generative control, our embeddings support localized action classification in multi-actor videos. Across both tasks, our approach improves controllability and outperforms global representations localized through cropping or post-hoc masking. Project page: https://compvis.github.io/WhatMoves

Controlling continuous, physically valid 3D surface evolutionis a fundamental challenge in computer vision and graphics. Current neu-ral representations typically rely on unstructured latent spaces or discreteiterations, lacking explicit geometric interpretability and struggling toproduce smooth deformations. While classical scale-space theory offersclear physical interpretability, its numerical irregularity isolates it frommodern deep learning. In this paper, we bridge this gap by establishing alearnable scale-space foundation. Our framework comprises three syner-gistic components: (1) the Variational Neighborhood Curvature (VNC)operator, an efficient, parallelizable, and scale-stable geometric metric,(2) the spacetime-balanced VNC-Flow algorithm, which translates ir-regular physical smoothing into structured deformation trajectories, and(3) a Controllable Variational Autoencoder (C-VAE) that learns thesetrajectories conditioned on normalized evolution time. Departing fromtraditional discrete iterations or unstructured latent interpolations, ourcontinuous neural surrogate induces a structured radial latent organiza-tion. This enables precise control over continuous shape abstraction andestablishes a structural prior for downstream applications.